Estimation of Link Speed Distribution from Probe Vehicle Data

Probes with GPS devices reveal useful information for traffic conditions, but the high level of noise and the sparsity of observations make it challenging to estimate speed distribution from the data collected. This paper proposes a Bayesian approach for estimating link speed distribution from GPS-equipped probe data. The key contribution of the study is a generic hierarchical Monte Carlo Markov chain algorithm for sampling from probe speed distribution, with Gaussian mixture models for probe speed clustering. The algorithm combines Gibbs sampling and Metropolis–Hastings sampling to improve convergence speed. A rigorous mathematical discussion is provided for the simulation approach. The algorithm is evaluated with synthetic data and real-world probe data and shows the feasibility of the approach. Results also confirm the computational advantages of the proposed algorithm and suggest its potential for real-time extension.